TerraMosaic Daily Digest: July 22, 2026
Daily Summary
Slope-failure studies converge on where water enters, how instability is expressed and what must be measured before failure. Sinkholes in Qinghai focus rainfall infiltration into loess and reproduce failure timing within the observed storm; in Chongqing, vegetation loss and a morphometric conditional-probability term identify rainfall-triggered hillslope flows with an AUC of 0.8846. Horne Lake supplies the longer view: 57 late-Holocene subaqueous landslides separate plausible earthquake and water-level triggers beyond the instrumental record.
Seismic-hazard evidence now links regional structure to transient fault motion. The CRESCENT Generation 0 model combines about two decades of receiver functions and ambient-noise measurements from roughly 2,300 stations into an uncertainty-resolved Cascadia velocity model. Five-minute GNSS records offshore Boso distinguish the onset, propagation and termination of four recurring slow-slip events, while receiver functions beneath Ecuador image the subducting Carnegie Ridge to 50-105 km inland. These constraints complement broadband simulation of the 2023 Kahramanmaras earthquake and site-specific studies of structural collapse, retrofit and ground response.
Across hazards, progress comes from retaining physical structure through the observation and inference chain. Fibre-optic seismology estimates that fractures occupy about 8% of sampled glacier ice; monitoring-constrained GPR resolves temperature and water-content contrasts within a thaw slump; and latent-wavelet deconvolution improves few-shot acoustic-impedance imaging without assuming a fixed seismic wavelet. Cross-scale satellite-UAV references, dual-angle radiative transfer and rotation-equivariant multispectral detection address different forms of geometric mismatch. Yet controlled perturbations of an atmospheric foundation model show that predictive skill can still coexist with chemically inconsistent internal behaviour.
Key Trends
Five shifts connect preferential flow, regional seismic structure, cascading hazards, physics-constrained inversion and transfer across scales.
- Preferential pathways replace uniform wetting assumptions: Sinkholes and karst hydrology determine where water reaches potential failure surfaces, while GPR-resolved hydrothermal contrasts show why point measurements miss local thaw-driven instability.
- Regional structure and transient slip are resolved together: Community velocity models, receiver-function imaging and high-rate GNSS connect three-dimensional plate structure to earthquake-wave propagation and evolving fault slip.
- Hazards are represented as connected sequences: Spatial storm networks, landslide-barrier-lake cascades and transboundary typhoon chains retain propagation paths and timing between nominally separate events.
- Hidden state is inferred with explicit physical constraints: Fibre-optic seismology, GPR inversion, acoustic deconvolution and layer-aware site-response models recover fracture, water, temperature and impedance fields that surface proxies cannot resolve.
- Transfer is tested across scale, geometry and geography: Cross-scale references, rotation-equivariant detection, strict spatial separation and station-disjoint tests distinguish reusable process information from interpolation or memorization.
Selected Papers
An open, uncertainty-resolved three-dimensional velocity model for Cascadia leads this issue, followed by high-rate GNSS constraints on recurring slow slip offshore Japan, preferential-seepage initiation of loess landslides, a late-Holocene archive of lake-floor failures and multi-source karst-landslide forecasting. Together, the studies connect subsurface structure, transient motion and process-specific monitoring across seismic, slope and cryospheric hazards.
1. The CRESCENT Generation 0 Cascadia Community Velocity Model: Constraints From the Joint Inversion of Teleseismic Receiver Functions and Ambient Noise Data
Core Problem: Cascadia ground-motion modelling lacks a regional three-dimensional velocity model with uncertainty estimates at scales relevant to basins and infrastructure.
Key Innovation: Joint Bayesian inversion of about 20 years of receiver functions and ambient-noise data from roughly 2,300 stations produces the open CRESCENT Generation 0 shear-wave velocity model.
2. Investigations on the initiation mechanism of rainfall-induced loess landslides driven by sinkhole-assisted seepage: a case study in Qinghai, China
Core Problem: Rainfall thresholds alone cannot explain why failure localizes rapidly in otherwise similar loess terrain.
Key Innovation: Field infiltration, strength testing and FLAC3D simulations identify sinkholes as preferential seepage conduits and reproduce initiation near the observed storm duration.
3. Late Holocene subaqueous landslide activity within Horne Lake, Vancouver Island, British Columbia, Canada
Core Problem: The recurrence and triggers of lake-floor slope failures are difficult to infer from short instrumental earthquake and water-level records.
Key Innovation: A 57-event subaqueous-landslide inventory extends the record through the late Holocene and tests drawdown against major regional earthquakes as competing triggers.
4. Multi-Source Environmental Information Fusion and Adaptive Deep Learning for Karst Landslide Displacement Prediction
Core Problem: Karst-landslide displacement reflects coupled rainfall, thermal and hydrological forcing that a single time series cannot represent.
Key Innovation: SAPSO-optimized decomposition and a GRU fuse GNSS, rainfall, soil-temperature and moisture observations, with reported next-day R-squared values above 0.95.
5. Study on the characteristics of debris flow movement in the transition section of inlet drainage channels
Core Problem: Abrupt channel transitions can amplify or redistribute debris-flow motion, but the separate effects of material density, contraction and slope are poorly constrained.
Key Innovation: Factorial flume tests rank channel slope as the dominant control and derive regression relations for the transition-zone mud-trace ratio.
6. Slope stability assessment in blocky rockmass using integrated approach of kinematics and block theory analysis
Core Problem: Conventional continuum stability analyses can miss removable blocks governed by discontinuity geometry in heavily jointed rock slopes.
Key Innovation: Kinematic analysis, block theory and finite elements jointly identify key blocks and minimum safe slope angles for the Tanahu hydropower excavations.
7. Research on selection and deployment of monitoring equipment based on local stability analysis model of slope
Core Problem: An acceptable global factor of safety can conceal local instability and lead to sparse or misplaced monitoring.
Key Innovation: A fault-tree local-stability model targets sensors to unstable sectors and explicitly balances equipment performance, redundancy and cost.
8. Performance Evaluation of a Tunnel–Slope System
Core Problem: Tunnel excavation and seismic loading can interact with an already fractured, rainfall-sensitive slope over a poorly defined spatial range.
Key Innovation: Three-dimensional finite-difference simulations locate the strongest tunnel-slope interaction within roughly four tunnel diameters.
9. A WebGIS-Based Platform for Sharing Earthquake Surface Rupture Data in Mainland China
Core Problem: Mainland China's surface-rupture observations are fragmented across publications and incompatible local data products.
Key Innovation: The MCSRD platform normalizes 632 source records for 72 earthquakes and exposes them through an interoperable PostGIS and OGC WebGIS stack.
10. Investigating the 2023 Mw 7.8 Kahramanmaraş Earthquake Ground Motion Through 3D Physics-Based Numerical Simulations
Core Problem: Regional observations alone cannot reveal how source, path and three-dimensional structure combined to produce the 2023 Kahramanmaras ground-motion field.
Key Innovation: A 325 by 355 km physics-based simulation is extended above 1 Hz with a neural broadband model and validated against recorded motions.
11. Beyond localized hotspots: hazard connectivity in spatially compound rainfall-coastal storm events
Core Problem: Local co-occurrence metrics miss compound storms whose rainfall and coastal components connect distant shoreline sectors.
Key Innovation: A 43-year Spanish record reveals two recurring connectivity modes and reframes compound hazard as a spatial network rather than a set of hotspots.
12. Deciphering volcanic activity: ground and satellite observations of Vulcano's La Fossa Crater and Baia di Levante (2018–2024)
Core Problem: At Vulcano, deformation, thermal and gas observations respond on different timescales, complicating interpretation of unrest escalation and decline.
Key Innovation: Six years of ground and satellite data resolve five phases and link delayed, near-synchronous signals to sealing and drainage of magmatic fluids.
13. A Comprehensive Hazard Index-Based Potential Flood Disaster Chain Identification Model in the Guanting Gorge Section of the Yongding River Basin
Core Problem: Assessing rainfall, landslides, barrier lakes and dam breaks independently loses the pathways through which a gorge-scale flood disaster escalates.
Key Innovation: A weighted comprehensive hazard index maps the full chain and reports an AUC of 0.901 for potential cascade identification.
14. Spatial heterogeneity of land subsidence in the california central valley: a model-based attribution perspective
Core Problem: Central Valley subsidence drivers vary spatially, so basin-wide averages cannot support differentiated groundwater controls.
Key Innovation: Explainable stacking, clustering and generalized additive models separate four regimes and identify nonlinear pumping, precipitation and compressible-layer thresholds.
15. High‐Rate GNSS Analysis of the Onset, Growth, and Termination of Slow Slip Events and Their Along‐Dip Variability Offshore the Boso Peninsula, Japan (2011–2024)
Core Problem: Daily GNSS solutions blur the onset and termination of recurring slow-slip events near the source region of potential Tokyo-area megathrust earthquakes.
Key Innovation: Five-minute GNSS, sparse change-point detection and Bayesian slip inversion resolve distinct nucleation, propagation and termination patterns for four Boso events.
16. Quantifying subsurface fracture damage in glaciers using fiber-optic seismology
Core Problem: Surface crevasse maps do not quantify the subsurface fracture volume that governs glacier instability, calving and ice-avalanche potential.
Key Innovation: Distributed acoustic sensing links seismic anisotropy and tensile icequakes to a consistent fracture estimate of about 8% of ice volume.
17. A framework for emergency susceptibility assessment of hillslope flows triggered by extreme rainfall based on morphometric conditional probability
Core Problem: Emergency mapping after extreme rain needs a rapid inventory and a terrain descriptor that captures where hillslope flows concentrate.
Key Innovation: Sentinel-2 vegetation loss, geomorphic filtering and a new slope-area conditional-probability factor raise random-forest susceptibility performance to AUC 0.8846.
18. Numerical modeling for soil slope stability analysis and mitigation measures along the Weyiza–Wulokode section towards Chencha, Southern Ethiopia
Core Problem: Highly saturated roadside slopes in southern Ethiopia require mitigation choices tied to quantified failure conditions rather than visual condition alone.
Key Innovation: Multiple limit-equilibrium formulations establish the unstable baseline and show that slope flattening and soil nailing raise factors of safety above 1.411.
19. Research trends and knowledge gaps in earthquake-induced landslides in Nepal: a bibliometric review
Core Problem: Earthquake-induced-landslide research in Nepal is dominated by short post-event studies centred on the 2015 Gorkha earthquake.
Key Innovation: A PRISMA-screened review of 130 papers maps methodological progress and identifies missing longitudinal, runout, ecological and western-Nepal evidence.
20. China–Korea Linked Typhoons: A transboundary sequential hazard framework for East Asian tropical cyclones
Core Problem: National typhoon catalogues obscure sequential impacts that cross the China-Korea boundary within a common storm window.
Key Innovation: Seventy years of tracks identify 36 linked events and quantify their shared paths and median six-day interregional timing.
21. Evaluating multi-hazard early warning systems in Oregon: an analysis of OR-Alert notification patterns
Core Problem: Warning-system capability is often assessed from plans rather than the alerts and delivery channels used during real events.
Key Innovation: The OR-Alert audit analyses 31,940 notifications and 11.1 million deliveries across hazard, place, channel and message pattern.
22. Integrated Hydro-Hazard Index (HHI) for Drought-Flood Risk Assessment: A Multi-Temporal Machine Learning Approach
Core Problem: Separate drought and flood indices cannot represent rapid shifts between opposite hydrological extremes in one risk framework.
Key Innovation: A multi-temporal hydro-hazard index and machine-learning classifier integrate both regimes, reaching a reported AUC of 0.967.
23. Towards a National Geospatial Digital Twin in Slovenia: Multi-Source Integration and a Flood Pilot for Hazard Management
Core Problem: National flood intelligence depends on combining long gauge records with rapid imagery inside a reusable geospatial architecture.
Key Innovation: The Slovenian pilot integrates 1954-2025 hydrology and 26 PlanetScope scenes, demonstrating an interoperable digital-twin workflow with F1 of 0.75.
24. Spatial Lock‐In and Reclamation Delay in Coal‐Subsidence Landscapes: Land Degradation Under Passive Water‐Retention Strategies
Core Problem: Passive water retention can stabilize mine-subsidence landscapes temporarily while delaying ecological and land-use recovery.
Key Innovation: A decadal trajectory analysis identifies spatial lock-in and quantifies how water-management choices alter reclamation timing and substrate condition.
25. Protected Agriculture Mapping as a Proxy for Groundwater Abstraction in Subsidence-affected Areas: A Remote Sensing Approach
Core Problem: Pumping records are often too sparse to explain where groundwater-driven agricultural subsidence is intensifying.
Key Innovation: Sentinel-2 protected-agriculture mapping is coupled to Sentinel-1 PSI deformation, testing greenhouse expansion as an observable abstraction proxy.
26. Quantifying the Surface Deformation of Pingos on the Alaskan North Slope Using Interferometric Synthetic Aperture Radar (InSAR)
Core Problem: Pingo activity is difficult to monitor consistently across remote Arctic terrain with field surveys alone.
Key Innovation: An eight-year InSAR time series measures 11 pingos and identifies net uplift of 1.49-11.35 cm in eight of them.
27. High‐Resolution 3‐D Lithospheric Structure of the Subducting Carnegie Ridge and the Ecuadorian Margin Imaged From Teleseismic Receiver Functions
Core Problem: The inland extent and crustal expression of the subducting Carnegie Ridge remain uncertain despite their influence on Ecuadorian earthquake behaviour.
Key Innovation: Dense receiver functions image an 18-20 km thick ridge crust that thins inland and trace its subduction roughly 50-105 km east of the coast.
28. Monkey King Bang: A Unified Scientific Multimodal Foundation Model
Core Problem: Scientific foundation models remain divided by discipline and often translate non-text data into interfaces that discard modality-specific structure.
Key Innovation: MKB uses one Transformer with modality-specific encoders and decoders for native scientific understanding and generation across molecular, biological, medical and weather tasks.
29. Toward Mechanistic Interpretability of an AI Foundation Model Fine-Tuned for Atmospheric Chemistry
Core Problem: High air-quality forecast skill does not establish that a weather foundation model has learned atmospheric chemistry.
Key Innovation: Controlled chemical perturbations and sparse-autoencoder probes show a first-order ozone response but chemically inconsistent species combinations and meteorology-dominated internal structure.
30. HyperImageNet: A Large-Scale High-Spatial Resolution Hyperspectral Imagery Classification Benchmark
Core Problem: Hyperspectral land-cover benchmarks remain too small and spatially entangled to test fine-grained transfer into unseen areas.
Key Innovation: HyperImageNet provides 26,084 raw 224-band patches, 138 classes, pixel labels, instance masks and a strict spatially separated open-environment protocol.
31. T-STAR: A Large-Scale Benchmark for Spatio-Temporal Panoptic Scene Graph Generation in Satellite Video
Core Problem: Object detection alone cannot represent identity, relation and temporal evolution in satellite video.
Key Innovation: T-STAR defines spatiotemporal panoptic scene-graph generation with 1.1 million masks and 3.8 million temporally bounded relationship triplets.
32. SEAS5-BCSD: a bias-corrected and downscaled global seasonal forecast reference dataset for 1981–2024
Core Problem: Bias and drift in raw seasonal forecasts limit drought, water and energy applications, especially across long hindcast periods.
Key Innovation: SEAS5-BCSD releases all ensemble members for 1981-2024 at 0.25 degrees and evaluates probabilistic temperature and precipitation skill through seven-month leads.
33. Westerly induced precipitation increase drives near-surface ground ice gain on the Tibetan Plateau despite warming
Core Problem: Regional warming does not explain field evidence of increasing near-surface ground ice across parts of the Tibetan Plateau.
Key Innovation: Boreholes, a physics-based permafrost model and satellite deformation attribute ice gain over 35% of plateau permafrost to precipitation strengthened by westerlies.
34. Seismic response and collapse mechanism of the historic Sirkeci Railway Station: field observations and nonlinear time-history analysis
Core Problem: Complex geometry and sparse documentation obscure how historic masonry stations progress from local cracking to collapse under near-field shaking.
Key Innovation: A survey-constrained nonlinear model localizes damage around central-hall windows and links concurrent in-plane and out-of-plane failure to roof collapse.
35. Monitoring-constrained GPR inversion reveals profile-scale hydrothermal heterogeneity in a retrogressive thaw slump
Core Problem: Point sensors cannot resolve the lateral temperature and water contrasts created by retrogressive thaw-slump disturbance.
Key Innovation: Monitoring-constrained GPR reconstructs two-dimensional hydrothermal fields with 0.219 degrees Celsius temperature RMSE and exposes profile-scale wet-zone contrasts.
36. Plate-type steel fuse for enhancing the seismic performance of steel storage rack structures
Core Problem: Conventional rack retrofits obstruct storage bays, while hook-type beam connections remain vulnerable to instability under strong shaking.
Key Innovation: A notched steel-plate fuse adds stiffness and dissipates energy at existing joints, moving a benchmark rack from global failure to the Collapse Prevention objective.
37. Prediction of ground response spectra using a convolutional U-Net with cross-attention to subsurface layers
Core Problem: Site proxies and fixed-depth velocity samples suppress the ordered stratigraphic context that controls period-dependent amplification.
Key Innovation: A conditional U-Net with cross-attention uses layer-wise inputs, improves in-domain spectra and reveals the need for lightweight site-specific updating under station transfer.
38. A Graph Neural Network approach to zero-shot Digital Twins
Core Problem: Predictive digital twins are commonly tied to one geometry and require retraining when the domain or boundary conditions change.
Key Innovation: A thermodynamics-informed graph network couples live visual geometry to a solver that enforces energy conservation and non-negative entropy production across unseen domains.
39. HypNO: A Graph-Based Neural Operator with Physics-Informed Message Passing for Hyperbolic Conservation Laws
Core Problem: Neural operators often smooth shocks or violate admissible transport when applied to hyperbolic conservation laws.
Key Innovation: HypNO performs upwind, entropy-aware message passing on finite-volume space-time graphs and preserves discontinuities across unseen initial conditions.
40. Joint Utilization of Geospatial and census proxies for Autoencoder-Assisted Downscaling (JUGAAD) of socioeconomic indicators in India
Core Problem: Household surveys report socioeconomic indicators too coarsely and infrequently for village-scale monitoring, while census and geospatial proxies differ in scale and noise.
Key Innovation: JUGAAD regularizes inputs on intermediate village-cluster tessellations and learns an autoencoder-assisted mapping from census and geospatial proxies to finer socioeconomic estimates.
41. Latent Variable-Mediated Cross-Learning for Few-Shot Acoustic Impedance Imaging
Core Problem: Few-shot acoustic-impedance inversion is unstable when the seismic wavelet is unknown and labelled well logs cover less than 1% of traces.
Key Innovation: A differentiable Tikhonov deconvolution operator estimates the latent wavelet during semi-supervised cross-learning without an auxiliary network or fixed prior.
42. GeoThreat: Transferable Targeted Adversarial Attacks on Large Vision-Language Models for Remote Sensing Image Interpretation
Core Problem: Robustness tests for remote-sensing vision-language models rarely ask whether a black-box attack can force a specific, transferable semantic error.
Key Innovation: GeoThreat aligns local discriminative cues with global scene semantics to construct targeted perturbations that transfer across models and remote-sensing interpretation tasks.
43. Loss Landscape Topology Reveals Why Simple Baselines are Competitive at 3D Point Cloud Segmentation Under Class Imbalance
Core Problem: Class-imbalance methods developed for two-dimensional images may not improve minority-class boundaries in three-dimensional point-cloud segmentation.
Key Innovation: An eleven-method comparison and loss-landscape analysis show standard cross-entropy remaining within 0.8-3.3 mIoU points of specialized alternatives across two imbalance regimes.
44. DAPM: UAV Monocular Depth Estimation from Any Height, Pitch, Roll and FOV
Core Problem: UAV depth models trained at fixed viewpoints fail when flight height, pitch, roll and field of view change together.
Key Innovation: DAPM derives an explicit camera-geometry representation and jointly estimates depth and pose across the new UAPD viewpoint-diverse aerial dataset.
45. Transformer-based Diffusion models for Hydrological Time Series Probabilistic Imputation and Forecasting
Core Problem: Sparse, drifting hydrological sensors leave gaps that deterministic interpolation cannot represent probabilistically.
Key Innovation: A Transformer diffusion model learns water quantity and quality across six sites and more than 15 years, sampling realistic distributions for both missing-data reconstruction and forecasting.
46. Three-dimensional particle morphology characterization in the field using a video-extracting and reconstruction (VER) photogrammetry
Core Problem: Field particle-shape measurements need three-dimensional accuracy without the cost and logistics of a laser scanner.
Key Innovation: Smartphone video extraction and automated photogrammetry reconstruct 41 of 50 field particles, with all successful morphology estimates below 9.853% error and mean errors below 3.546%.
47. Rivers in motion: a wavelet-based remote sensing approach to spatio-temporal hydro-geomorphic analysis
Core Problem: Channel-centreline maps show where a river moved but do not isolate the spatial scale and intensity of meander change.
Key Innovation: NDWI-derived Aras River positions and continuous wavelet analysis quantify 1987-2019 migration of 250 m to 1 km and test its significance with Monte Carlo uncertainty.
48. Dynamic response of underground structures under coupled train loading and obliquely incident SV-Wave
Core Problem: Metro tunnels can experience train vibration and obliquely incident seismic waves simultaneously, a loading combination omitted from many response models.
Key Innovation: A validated three-dimensional soil-tunnel-track model introduces viscoelastic boundaries and equivalent oblique SV-wave loading to resolve the coupled response.
49. U-CFR: Uncertainty-Guided Cascade Forward Refinement for Interactive Segmentation
Core Problem: Interactive segmentation still requires repeated corrective clicks because refinement remains passive between user inputs.
Key Innovation: U-CFR converts boundary uncertainty, contour gradients and edge predictions into internal pseudo-clicks that target the most ambiguous regions after each interaction.
50. Spectral-Spatial Synergistic Guided Network for Hyperspectral Salient Object Detection
Core Problem: Hyperspectral saliency models confuse illumination-driven spectral variation with material differences that define the target.
Key Innovation: S3GNet couples robust spectral modelling, cross-stream spatial guidance and multiscale refinement in a lightweight closed-loop architecture.
51. Geo3R: Mitigating Spatial Reasoning Hallucination in Multimodal Large Language Models
Core Problem: Multimodal language models infer spatial relations from two-dimensional features that can contradict the underlying three-dimensional scene.
Key Innovation: Geo3R reconstructs and reasons over explicit spatial structure at inference time, reducing hallucinations across 18 tasks without additional model training.
52. DINOde: Continuous Vision-Text Alignment for Open-Vocabulary Semantic Segmentation
Core Problem: DINOv3 supplies strong spatial features but lacks the text alignment required for open-vocabulary segmentation.
Key Innovation: DINOde evolves CLIP text embeddings through continuous ODE flows toward the DINO visual manifold while injecting image-level context into pixel classification.
53. A real-time RGB-D perception pipeline for autonomous impact hammers in mining: self-filtering, rock segmentation and rock-breaking poses generation
Core Problem: Underground rock breakers need collision-free tool poses and a current three-dimensional workspace model without relying on teleoperation.
Key Innovation: An embedded RGB-D pipeline removes robot geometry, segments rocks, reconstructs the workspace and generates feasible impact poses in real time.
54. Structure-Preserving Physics-Informed Neural Network for the Korteweg--de Vries (KdV) Equation
Core Problem: Standard physics-informed networks can drift from conserved mass and energy during long integrations of nonlinear dispersive waves.
Key Innovation: The proposed KdV solver embeds mass and Hamiltonian invariants directly in training to maintain physically consistent, energy-stable evolution.
55. S-Agent: Spatial Tool-Use Elicits Reasoning for Spatial Intelligence
Core Problem: Spatial agents usually reason from isolated frames and lose the scene state accumulated across views and time.
Key Innovation: S-Agent treats tool use as continuous spatiotemporal evidence accumulation, maintaining a scene-centred representation across multi-view images and video.
56. Fully Rotation-Equivariant Spectral-Spatial Learning for Multispectral Object Detection
Core Problem: Multispectral detectors treat bands discretely and lose geometric consistency when object orientation and spectral reliability vary across scales.
Key Innovation: FressDet combines continuous spectral warping, reliability-weighted fusion and rotation-equivariant prediction, reporting state-of-the-art results with 93% fewer parameters.
57. Edge-preserving and uncertainty-guided domain adaptation for weakly supervised rooftop photovoltaic mapping
Core Problem: Rooftop-photovoltaic mapping degrades across cities and sensors, while dense target-domain labels remain too expensive.
Key Innovation: EPUDA transfers from dense source labels to sparse target points using edge preservation and uncertainty-guided adaptation for low-cost cross-domain segmentation.
58. Self-supervised spectral unmixing using scene-aware deep priors
Core Problem: Separating hyperspectral denoising, endmember discovery and abundance estimation compounds error under noise and spectral variability.
Key Innovation: A scene-aware self-supervised prior optimizes denoising and unmixing jointly while regularizing spatial and spectral fidelity without external labels.
59. A residual-based approach to downscale all-sky LST using synthetic clear-sky priors
Core Problem: All-sky land-surface-temperature products remain too coarse for local applications, and clear-sky downscaling assumptions fail beneath clouds.
Key Innovation: A synthetic clear-sky prior defines the fine spatial pattern, while residual adjustment transfers observed all-sky anomalies from roughly 5 km to 1 km.
60. RealVDeblur: One-Step Diffusion for Generalizable Real-World Video Deblurring
Core Problem: Video-deblurring systems trained on narrow synthetic kernels do not generalize to mixed camera and object motion in real captures.
Key Innovation: RealVDeblur combines physically grounded blur synthesis from 3D Gaussian scenes and high-frame-rate video with a one-step generative restoration model.
61. HalluScope: Fine-grained Hallucination Diagnosis for Multimodal Large Language Models
Core Problem: Binary hallucination detection cannot distinguish visual, textual and commonsense failure modes or guide targeted correction.
Key Innovation: HalluScope defines a fine-grained taxonomy, builds a 30,000-example diagnostic corpus and trains joint detection-classification models across five task categories.
62. DTIF: Robust Loop Closure Detection via Delaunay Triangle Topology in Complex Forests
Core Problem: Sparse low-cost LiDAR and repetitive trunks make initialization-free map registration unreliable in GNSS-denied forests.
Key Innovation: DTIF encodes trunks with Delaunay topology and uses topology-weighted pose estimation for lightweight loop closure and global registration.
63. QATMA: Quantization-Aware Training with Multimodal Alignment for Open-Vocabulary Object Detection
Core Problem: Very low-bit open-vocabulary detectors lose both region-text and region-region alignment during quantization.
Key Innovation: QATMA combines curriculum quantization with multimodal alignment losses, improving reported low-bit performance by up to 4.3 AP on LVIS and 7.6 AP on COCO.
64. A GIS-based multi-criteria geospatial workflow for mapping the hazard, exposure and risk of explosive remnants of war in twentieth-century conflict landscapes: a case study from northeastern Italy
Core Problem: Explosive-remnant risk maps often separate wartime evidence from present exposure and provide little independent validation.
Key Innovation: A reproducible GIS workflow integrates archives, aerial imagery, crater mapping, clearance records and current exposure, with 73.8% of 339 discoveries in the highest hazard classes.
65. Toward accurate disease diagnosis of fruit trees: Integrating a novel dual-angle vegetation index with canopy vertical distribution knowledge
Core Problem: Single-angle canopy observations confound disease severity with its vertical distribution.
Key Innovation: A dual-angle vegetation index and three-dimensional radiative-transfer model retrieve vertical disease position before constraining severity estimation from UAV observations.
66. BLB_TVI: A hyperspectral triangular index for satellite-scale characterization of rice bacterial leaf blight severity
Core Problem: Field disease scores and UAV observations do not align directly with the support of a 10 m satellite pixel.
Key Innovation: A cross-scale reference links field, UAV and Zhuhai-1 data to derive an interpretable three-band index that remains predictive across year and site.
67. Machine learning correction of C2-Net for chlorophyll-a retrieval in optically complex small reservoirs
Core Problem: Atmospheric-correction processors preserve useful signal in small reservoirs but introduce large, systematic chlorophyll-a bias.
Key Innovation: Machine-learning correction retains the physics-based processor outputs while reducing retrieval RMSE by up to 35% across 32 reservoirs.
68. Selection of optimal spatial window size for winter wheat LAI retrieval based on UAV multispectral imagery and the PROSAIL model
Core Problem: UAV retrieval accuracy depends on the spatial support used to average reflectance, yet hybrid physical-data models rarely optimize that scale.
Key Innovation: PROSAIL simulations and multispectral observations identify growth-stage-specific windows and improve cross-year leaf-area-index retrieval.